Imagine a community arts center losing access to one of its rooms. An AI recommends canceling the weekday music class because it has the fewest attendees.
The director knows some regulars cannot attend at another time. Keeping that class may matter more to the center's purpose than filling every seat. The AI can help analyze that consideration too, if someone supplies it. But producing an argument is not the same as having authority to settle what the center is for.
That distinction is central to how I think about AI.
We should use capable tools to explore options and do work. We should also be clear about who chooses the goals, who may change them, and who has to explain a consequential decision to the people affected.
This is an operating principle, not a claim that the law automatically makes every person in an AI chain liable for everything it does. Legal responsibility depends on the circumstances. Nor does my principle prove there will be a paid supervision job for every worker. One person may oversee work that once took many people.
The NIST AI Risk Management Framework Playbook emphasizes clear roles, responsibilities, and decision-making structures. That is practical guidance for organizing AI work, not a promise about how many jobs will remain.
The strongest objection is that humans often approve recommendations without understanding them. People make mistakes, too. Putting a human name next to a machine's answer does not make it wise, fair, or accountable.
For the arts center, useful oversight would require someone with time to examine the recommendation and authority to reject it. They would need to hear what the affected people think. If the director cannot change the outcome, the signature is just decoration.
The same question applies to smaller, everyday uses. May an assistant draft a customer reply, or send it? May it suggest a purchase, or spend the money? May it identify an old file, or delete it? These are different grants of authority, even when one tool can do all of them.
Before using an assistant, I would put four things in plain language: the result we want, the boundaries it must respect, the decisions a person must approve, and what happens when it is uncertain. Those choices deserve as much attention as whether the tool can complete the task.
As AI gets more capable, I expect those choices to become more important, not less. The goal is not to keep a person busy pressing yes. It is to preserve a meaningful ability to decide what should happen and to stop something that should not.
Before your next AI project, ask a simple question: who has both the time and the authority to say no?
AI-assisted editing. Published with Rob Price's approval.